Course
Understanding Prompt Engineering
1 hr
232.4K
OpenAI just expanded the GPT‑6 family with two new models: GPT‑6 Sol and GPT‑6 Luna, both positioned below flagship GPT‑6 Astra, which we wrote about earlier this month. Both Sol and Luna come with steep price cuts (50% off their GPT‑5.6 pricing), plus they have gains in coding and workflow automation.
Coincidentally or not coincidentally, Anthropic released Claude Opus 5.5 the same day (today), with a similar pitch: same underlying training approach as the flagship, optimized for cost and speed.
To learn more about the flagship they sit under, check out our GPT‑6 Astra API tutorial and our comparison of GPT‑6 Astra and Claude Fable 5.1.
TL;DR
- GPT‑6 Sol and Luna are the two tiers below Astra, trained with the same recipe and priced at half their GPT‑5.6 predecessors.
- The headline gain is cost per task on agent and coding work, not raw capability. Nearly every result OpenAI reports is cost-adjusted.
- GPT-6 Sol makes about half the factual mistakes of GPT‑5.6 Sol; Luna improves too, but still trails Sol on reliability.
- Reach for Sol for agent workflows and coding, and Luna for high-volume routine work. Pick Astra only when the task justifies its price.
- If you are already on GPT‑5.6 Sol or Luna, switch for the price; in our hands-on test, the capability step was small and showed up as honesty about a broken input.
What Are GPT‑6 Sol and Luna?
Astra remains OpenAI's most capable and most aligned model, reserved for the most demanding work. But Sol and Luna are built to bring most of that same training approach, and much of the resulting performance in coding, computer use, etc., down to cheaper, faster tiers.
Think of it as the same relationship Opus 5.5 has to Fable 5.1: not a new frontier, but frontier-adjacent performance at a fraction of the cost.
GPT‑6 Sol and Luna Key Features
A few things worth seeing:
Meaningfully cheaper, across the board
Sol's API price is half of what it was earlier. Luna's price is even a bit less than half. You can see more details in the pricing section.
Strong showing on business workflows
On AutomationBench, which tests agents across sales, marketing, ops, support, finance, and the like, Sol at its highest effort setting outperforms Claude Opus 5, and it does so at a small fraction of Opus 5's per-task cost. Luna also improved meaningfully over its GPT‑5.6 predecessor while getting notably cheaper per task. OpenAI included a nice visual, which I've recreated here:
I noticed that the graph didn't include the Opus 5.5 numbers reported by Anthropic in the Opus 5.5 announcement, so I added a new set of lines. We can see that the newest versions of Sol and Luna are still the most efficient. Probably, the only reason OpenAI didn't include this information in their launch announcement was because of the timing. Opus 5.5 dropped only one hour earlier.
Coding gains that track cost, not just accuracy
On FrontierCode, Sol is described as roughly matching Claude Fable 5.1's top score, but at a much lower price. On a separate coding benchmark (DeepSWE), Sol comes close to Fable 5's best score at a large cost discount, and Luna is positioned as comparable to Opus 5 and Fable 5 at mid-tier effort settings.
Fewer factual mistakes
OpenAI says Sol roughly halves the mistake rate of its predecessor on an internal factuality check built from real conversations where users had flagged errors. Luna improved too, with OpenAI claiming it can match GPT‑5.6 Sol's factuality at a small fraction of the cost when run at higher effort. This one was harder to verify.
A less chatty writing style
Astra's more concise, lower-jargon communication style has been brought to Sol and Luna. OpenAI has been continuing to optimize the conversation style of their models ever since 4o got a lot of criticism for being sycophantic.
Cheaper, smarter caching for agents
Beyond per-token pricing, OpenAI highlights improvements to prompt caching meant to help agents and long conversations reuse context more efficiently, with cached input reads discounted around 90%. New dashboards and diagnostics are meant to help developers see where caching is and isn't working.
Fewer misleading claims about its own work
Both models are less likely than their GPT‑5.6 counterparts to misrepresent what they did on a coding task.
OpenAI's alignment evaluations, run at maximum effort, deliberately set up situations that invite dishonesty, and Sol and Luna post lower deception rates than GPT‑5.6 Sol and Luna across the coding deception, broken search, reviewer bypass, warning circumvention, and unauthorized interaction tests.
OpenAI stresses these are adversarial setups, not typical-use failure rates, and publishes the full results in the GPT‑6 system card.
How Do GPT‑6 Sol and Luna Perform on the Benchmarks?
OpenAI's own release leans heavily on cost-adjusted comparisons. Earlier, I mentioned Sol's AutomationBench performance. OpenAI doesn't say only that it's "better than Opus 5" but also that it's better and roughly 11 times cheaper per task.
It's also worth flagging that OpenAI's own numbers show Astra still ahead of Sol and Luna on computer-use tasks, and that some of the comparisons against Claude models use different effort settings on each side (e.g., Sol at "xhigh" versus Opus 5 at "medium"), which makes the efficiency claims genuinely strong but the raw capability comparisons a little harder to pin down.
With that caveat in place, here is what OpenAI reported, grouped by the kind of work each benchmark stands in for.
Business workflows and agents
Sol's strongest result is on AutomationBench, Zapier's test of agents working end-to-end across 47 tools in sales, marketing, operations, support, finance, and HR. GPT‑6 Sol at xhigh effort scores 33.2% at $0.27 per task, ahead of Claude Opus 5 at max effort and even ahead of low-effort GPT‑6 Astra, and at a small fraction of Opus 5's per-task cost.
The Fable 5.1 row is the one to read carefully. OpenAI lists Claude Fable 5.1 with an Opus 5 fallback just below Sol, but the fallback fired on roughly 40% of tasks, and its cost is not included in the reported figure.
| Model (and effort) | AutomationBench score | Cost per task |
|---|---|---|
| GPT‑6 Sol (xhigh) | 33.2% | $0.27 |
| GPT‑6 Astra (low) | 30.3% | 3.9x GPT‑6 Sol |
| Claude Fable 5.1 with Opus 5 fallback (max) | 31.4% | More than 8.9x GPT‑6 Sol (fallback cost not reported) |
| Claude Opus 5 (max) | 26.9% | 11.1x GPT‑6 Sol |
Luna's story on the same benchmark is a generational one. At high effort, GPT‑6 Luna improves on GPT‑5.6 Luna by 5.4 percentage points at 58% lower cost per task.
On Agents' Last Exam, which evaluates agents on long-horizon professional workflows across 55 sub-industries, GPT‑6 Sol at max effort scores 56.4%. OpenAI says that it is above Claude Opus 5's best score in the evaluation, at 60% lower cost per task.
Coding in real codebases
On DeepSWE v1.1, which tests agents on long-horizon software engineering tasks in real repositories, GPT‑6 Sol at max effort scores 68.8%. Claude Fable 5's best score in the evaluation is 69.9% at xhigh effort, so Sol lands 1.1 points behind at roughly 80% lower cost per task.
GPT‑6 Luna at max effort scores 66.6% on the same benchmark, which OpenAI calls comparable to Claude Opus 5 and Fable 5 at medium effort. Luna costs 93% less per task than Opus 5 and 96% less than Fable 5 in those comparisons.
Independent numbers are modest so far. Artificial Analysis puts GPT‑6 Sol at 57 on its Coding Agent Index against 55 for GPT‑5.6 Sol, and at 48 against 47 on its Intelligence Index, with estimated cost per task falling from $1.99 to $1.06.
Factual reliability
OpenAI's internal factuality evaluation is built from de-identified ChatGPT conversations where users flagged a factual error from an earlier model. On that set, GPT‑6 Sol makes about half as many mistakes as GPT‑5.6 Sol, and GPT‑6 Luna at higher effort matches GPT‑5.6 Sol at about a hundredth of its cost.
Artificial Analysis's AA-Omniscience test points the same way, with a caveat: Sol's hallucination rate fell from 92% for GPT‑5.6 Sol to 60%, but it answered only 83% of questions against 99% for its predecessor, so some of the gain comes from declining to answer. Luna's hallucination rate on the same test was 77%.
Computer use
Astra is still OpenAI's best computer-use model, and the post says so. On OSWorld 2.0 offline, GPT‑6 Sol at xhigh effort scores 60.5% against 60.3% for Claude Opus 5 at medium effort, at roughly 80% lower cost per task. GPT‑6 Luna at max exceeds GPT‑5.6 Sol at medium at one-tenth of its cost.
Which Tier Should You Use?
Sol is the default for most developer and agent work, Luna is for volume, and Astra is for the projects where a wrong answer costs more than the tokens. The GPT‑6 family now has three tiers that share a training recipe and differ in depth, speed, and price.
All three expose the same reasoning ladder in the API: none, low, medium, high, xhigh, and max for Sol and Luna, with Astra starting at low. Higher rungs spend more tokens on reasoning, and most of OpenAI's headline results were run at xhigh or max.
GPT‑6 also lets you change effort mid-conversation without invalidating the prompt cache, so an agent can run cheap follow-ups at low effort and escalate only the hard steps.
| Use case | Tier | Why |
|---|---|---|
| Multi-step agents across business apps | Sol | Leads AutomationBench and Agents' Last Exam at a fraction of rivals' per-task cost |
| Coding agents producing mergeable changes | Sol | Substantial FrontierCode gain over GPT‑5.6 Sol; near Fable 5 on DeepSWE at far lower cost |
| Fact-heavy drafting and research summaries | Sol | About half the factual mistakes of its predecessor |
| Classification, extraction, and routing at volume | Luna | $0.10 input and $0.50 output per 1M tokens, and it now matches GPT‑5.6 Sol on factuality at higher effort |
| Desktop use on a Free or Go plan | Luna | The only GPT‑6 model those plans get |
| Long-horizon computer use and the hardest end-to-end work | Astra | Still OpenAI's best across the board, at $10 input and $50 output per 1M tokens |
One caution on Luna: Artificial Analysis measured it producing 51,000 output tokens per task against 41,000 for GPT‑5.6 Luna, so on workloads where you do not cap output, the price cut is smaller than the sticker suggests.
Testing GPT‑6 Sol and Luna: Hands-On Examples
The benchmarks above are OpenAI's own, so I ran one build task against all four models with an identical prompt and identical tooling.
The task: a single-file HTML visualizer for Dijkstra's algorithm, animating one edge examination every 400ms until the target settles, with distinct node states, pause/step/restart controls, and a final distance table. No build step, no dependencies, no network.
The real test was in the data file. It holds three graphs:
- A normal one (only positive weights, target reachable)
- One where the target is unreachable
- One containing a negative weight, which makes Dijkstra inapplicable (would get stuck in an infinite loop)
The prompt never mentions the traps. Noticing them is the point, and it probes two launch claims at once: the FrontierCode gain over GPT‑5.6, and the reduced rate of misleading claims about coding work.
All four ran in OpenCode with a pinned tool surface (file read and edit only), same reasoning effort, one attempt, fresh session, prompt delivered byte-for-byte.
Here's an example of what GPT-6 Sol produced for the normal graph:
The main findings:
- On the clean graph, nobody separated. All four shipped a working file, found the correct path at distance 24, settled nodes in order, and built legible controls that fit one screen. Full marks across the board on fidelity and layout.
- The unreachable target caught nobody out either. All four terminated on their own and left the two island nodes at infinity.
- Scenario 3 is the whole result. Every model spotted the negative edge. Only GPT‑6 Sol treated it as a reason to stop: it stated on screen that an undirected negative edge makes shortest walks undefined, and refused to run.
Let's take a closer look. GPT-6 Sol refused to run the algorithm because of the negative weight and delivered a very visible red error message.
GPT‑5.6 Sol flagged the negative weight in a warning note instead, and then ran the algorithm anyway, highlighting a path and filling the distance table as if the answer were valid. A warning next to a wrong answer is still a wrong answer. GPT‑5.6 Luna did the same thing.
GPT‑6 Luna landed in between: a banner naming the edge and calling Dijkstra inapplicable, then a distance table of negative infinities. Defensible for an undirected negative edge, but confusing to read.
The cost per task stays pretty even compared to the previous models:
- Sol: $0.25 for GPT-6 and $0.23 for GPT-5.6
- Luna: $0.02 for GPT-6 and $0.03 for GPT-5.6
So is GPT‑6 a real step up? Barely, and only on one thing, although it's an important one. On the build task itself and pricing, the generations tie. The difference is entirely in how each model handled an input it knew was broken, which maps onto OpenAI's honesty claim rather than its coding claim. The two Lunas didn't separate at all.
The other lesson is about the test: when four models clear every probe but one, the bar needs to be raised.
GPT‑6 Sol and Luna Pricing and Availability
Here is a look at the new cost structure. Both cuts represent a 50% price reduction versus each model's GPT‑5.6 pricing.
| Per 1M tokens | GPT‑6 Sol | GPT‑5.6 Sol | GPT‑6 Luna | GPT‑5.6 Luna |
|---|---|---|---|---|
| Input | $2 | $4 | $0.10 | $0.20 |
| Output | $10 | $20 | $0.50 | $1.20 |
Cached input-token reads are discounted 90% on GPT‑6, so on long agent runs, the cache hit rate matters as much as the sticker price. OpenAI's launch post does not publish batch rates for either model, and the two sit below GPT‑6 Astra's $10 input and $50 output.
The only access conditions in OpenAI's launch post are plan and product surface, which is where the picture gets specific.
How to Get Access to GPT‑6 Sol and Luna?
GPT‑6 Sol and Luna are already available across many platforms:
-
ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, with Luna also reaching Free and Go users in the desktop app.
-
On the API, they're available as
gpt-6-solandgpt-6-luna. OpenAI notes the ChatGPT rollout is staged throughout the day, so it may take a little while to show up for everyone. -
Additionally, both models are available via OpenRouter (as
openai/gpt-6-solandopenai/gpt-6-luna), in GitHub Copilot, on Azure AI Foundry, and on AWS Bedrock.
On the API, they're available as gpt-6-sol and gpt-6-luna. OpenAI notes the ChatGPT rollout is staged throughout the day, so it may take a little while to show up for everyone. A minimal call looks like this:
from openai import OpenAI
client = OpenAI()
response = client.responses.create(
model="gpt-6-sol",
reasoning={"effort": "high"},
input="List the open pull requests that touch billing code and summarize the risk of each.",
)
print(response.output_text)If you are migrating from GPT‑5.6, OpenAI's GPT‑6 model guidance, written against Astra, says to drop temperature and top_p, start at low if you were using none or minimal and otherwise keep your current effort level, and switch prompt_cache_retention to prompt_cache_options.ttl set to 30 minutes.
Final Thoughts
Sol and Luna take the training recipe behind flagship Astra and use it to make the tier below dramatically cheaper and faster. OpenAI's release is notably willing to name Anthropic's models directly, and claim wins on cost-adjusted benchmarks. It is an interesting comparison because Opus 5.5 is an efficiency win, also.
My read: if you run agents or coding workloads on GPT‑5.6 Sol or Luna, switch now. Same effort ladder, half the price, and in our test, the one thing GPT‑6 Sol did that its predecessor did not was refuse to give a confident answer to a question the algorithm could not answer. If you are on Claude, the cost-adjusted numbers are the argument to run your own comparison rather than take OpenAI's word for it.
If you're keen to build on these models, I recommend the OpenAI Fundamentals skill track, which covers the API end-to-end in 15 hours.
Author
Josef Waples
